add more comments on handling small scale handling; remove redundant …

…if (pytorch#232)

Summary:
Pull Request resolved: pytorch#232

As title

Reviewed By: dskhudia

Differential Revision: D19231440

fbshipit-source-id: fe4db92bc6b5dd2822dbdac75f29b779d80cee65

src/PackAWithQuantRowOffset.cc

@@ -46,7 +46,7 @@ PackAWithQuantRowOffset<T, accT>::PackAWithQuantRowOffset(
   if (!cpuinfo_initialize()) {
     throw std::runtime_error("Failed to initialize cpuinfo!");
   }
-  if (scale_ == 0.0f || std::isinf(1.0f / scale_)) {
+  if (scale_ == 0.0f) {
     throw std::runtime_error("scale cannot be zero");
   }
   if (std::isinf(1.0f / scale_)) {

src/QuantUtils.cc

@@ -42,7 +42,9 @@ TensorQuantizationParams ChooseQuantizationParams(
   // final number to reflect the actual number used during quantization.
   float scale = (static_cast<double>(max) - min) / (qmax - qmin);
   // If scale is 0 or too small so its reciprocal is infinity, we arbitrary
-  // adjust the scale to 0.1
+  // adjust the scale to 0.1 . We want to avoid scale's reciprocal being infinity
+  // because some of fbgemm code pre-computes scale's reciprocal to do
+  // multiplication instead of division in the time critical part of code.
   if (scale == 0.0f || isinf(1.0f / scale)) {
     scale = 0.1;
   }